multiagent systems...multiagent system? coherence : how well does the system behave as a unit ?...
TRANSCRIPT
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Multiagent Systems7. Working Together: Methods for Coordination
B. Nebel, C. Becker-Asano, S. Wöl�
Albert-Ludwigs-Universität Freiburg
June 4, 2014
1 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Motivation
2 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Cooperative Distributed Problem Solving (CDPS)
Basic question:
How can a loosely coupled network of problem solvers worktogether to solve problems that are beyond their individualcapabilities? (Durfee et al., 1989, after Wooldridge, 2009)
Case one:Agents are benevolent, share a common goal and no potentialof con�ict between them.⇒ greatly simpli�es designer's task
Case two:Societies of self-interested agents don't share a common goal.⇒ designed by di�erent individuals or organizations to reachgoals through cooperation (which is nevertheless inevitable)
Second case is more challenging and interesting!3 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
CDPS versus parallel problem solving
How is CDPS di�erent from parallel problem solving (PPS)?
PPS exploits parallelism to solve problems
Computational components are processors, not agents
Central node decomposes overall problem intosubcomponents
Processors independently provide subsolutions
Subsolutions are assembled into overall solution by centralnode
PPS was synonym for CDPS in �early days of multiagentsystems�, two �elds are now quite separate.
4 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
How to evaluate a MAS
What criteria can be used to evaluate the �success� of amultiagent system?
Coherence: How well does the system �behave as a unit�?Criteria: solution quality, e�ciency of resource usage,conceptual clarity of operation, performance in thepresence of uncertainty/failure
Coordination: the degree to which agents can avoidine�ectual activitiesCriteria: amount of communication, task sharing, resultsharing
5 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Main issues in CDPS
Main issues to be addressed in CDPS include:
How can a problem be divided into smaller tasks fordistribution among agents?
How can problem solution be e�ectively synthesized
from subproblem results?
How to optimize overall problem-solving activities ofagents so as to produce solution to maximize coherencemetric?
What techniques can be used to coordinate activity ofagents, thus avoiding destructive interactions, andmaximizing e�ectiveness?
6 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
The three stages of CDPS
(Wooldridge, 2009, p. 154)
7 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Task sharing
8 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Task sharing
�How can a problem be divided into smaller tasks fordistribution among agents?�⇒ Task-sharing
(Wooldridge, 2009, p. 156)A task is decomposed into subproblems that are allocated toagents.
9 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Task sharing in the Contract Net
The Contract Net (CNET) protocol (Smith 1977, 1980):
high-level protocol for achieving e�cient cooperationthrough task sharingbasic metaphor: contracting
10 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Contract Net protocol
One of the oldest, most widely used agent interactionprotocols
A manager agent announces one or several tasks, agentsplace bids for performing them
Task is assigned by manager according to evaluationfunction applied to agents' bids (e.g., choose cheapestagent)
Idea of exploiting local cost function (agents' privateknowledge) for distributed optimal task allocation
Even in purely cooperative settings, decentralization canimprove global performance
A typical example of �how it can make sense to agentify asystem�
Successfully applied to di�erent domains (e.g. transportlogistics)
11 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Contract-net protocolAgent-Based Systems
Contract-net protocol
Initiator Participant
cfp
refuse
not−understood
propose
reject−proposal
accept−proposal
failure
inform−ref
inform−done
24 / 25(FIPA Contract Net Interaction Protocol Speci�cation, 2000) 12 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Marginal cost of task
How should a potential contractor decide whether or not to bidfor a task?Idea: marginal cost for task (Sandholm 1999)
Let t be the time, i be the contractor, τ ti be the taskscheduled for i at time t
The endowment of contractor i is ei and an announcementts is made with τ(ts) denoting the task speci�ed by ts
Let cti(τ) denote the cost to agent i of carrying out the setof task τ at time t
Then, the marginal cost of carrying out tasks τ is denoted by:
µi(τ(ts) | τ ti ) = ci(τ(ts) ∪ τ ti )− ci(τ ti )
⇒ di�erence between cost of carrying out old plus new tasksand only carrying out previously agreed tasks
13 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Blackboard architecture
Inspired by human problem solving in a team:�Imagine a group of human specialists seated next to a largeblackboard. The specialists are working cooperatively tosolve a problem, using the blackboard as the workplace fordeveloping the solution.Problem solving begins when the problem and initial data arewritten onto the blackboard. The specialists watch theblackboard, looking for an opportunity to apply theirexpertise to the developing solution. When a specialist �ndssu�cient information to make a contribution, she records thecontribution on the blackboard, hopefully enabling otherspecialists to apply their expertise. This process of addingcontributions to the blackboard continues until the problemhas been solved.� (Corkill, 1991)
14 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Blackboard metaphor
15 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Blackboard architecture characteristics
Eight characteristics of blackboard architectures (Corkill, 1991):
Independence of expertise (I think, therefore I am.)
Diversity in problem-solving techniques (I don't thinklike you do.)
Flexible representation of blackboard information (Ifyou can draw it, I can use it.)
Common interaction language (What'd you say?)
Positioning metrics (You could look it up.)
Event-based activation (Is anybody there?)
Need for control (It's my turn.)
Incremental solution generation (Step by step. . . )
16 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Basic Blackboard System
A basic blackboard system (Corkill, 1991):
KS: static Knowledge SourceKS Activations: Combination of KS knowledge and speci�c�triggering context� ⇒ the proactive �agent�
17 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
Blackboard architecture components
The simple blackboard system consists of two majorcomponents:
1 The blackboard as global structure available to all KSs,serves as:
a community memory of raw input data; partial solutions,alternatives, and �nal solutions; and control informationa communication medium and bu�era KS trigger mechanism
2 Control component as explicit control mechanism,chooses course of action based on state of blackboard andset of triggered KSs
Incremental reasoning style with the solution being built onestep at a time. At each step:
execute any triggered KS
choose a di�erent focus of attention, on the basis of thestate of the solution
18 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Contract Netprotocol
BlackboardArchitecture
Coordination
Models forCoordination
When to use Blackboard Problem-Solving Approach
Situations suitable for the blackboard approach (Corkill, 1991):
diverse, specialized knowledge representations
as an integration framework for heterogeneousproblem-solving representations and expertise (e.g.,diagnostic systems)
software development in teams of independent developers(i.e. as software-engineering metaphor)
When �uncertain knowledge or limited data inhibitsabsolute determination of a solution�
To realize multilevel reasoning or �exible, dynamiccontrol of problem-solving activities
The �Cognitive Agent Architecture � Cougaar� (DARPA) isbased on a blackboard architecture, but not FIPA-compliant. Itis used for the development of applications in military logistics.
19 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Coordination
20 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Methods for coordination
Coordination
Coordination is the process of managing inter-dependenciesbetween agents' activities
Coordination is a special case of interaction in whichagents are aware how they depend on other agents andattempt to adjust their actions appropriately.
Actually this only covers agent-based coordination, butthere can also be centralised mechanisms
In contrast to cooperation, coordination is also necessary innon-cooperative systems (unless agents ignore each other)
21 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Coordination relationship typology
A typology in the context of multiagent planning is provided by(von Martial, 1990):
22 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
von Martial's typology
von Martial's typology distinguishes:
Positive relationships: relationships between two agents'plans for which bene�t will be derived for at least oneagent if plans are combined
Requests: explicitly asking for help with own activities
Non-requested: pareto-like implicit relationships
action equality relationships: su�cient if one agentperforms action both agents needconsequence relationships: side e�ects of agent's planachieve other's goalsfavor relationships: side e�ects of agent's plan make goalachievement for other agent easier
Basic di�erence to traditional computer systems⇒ coordination is achieved at run time rather than designtime
23 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Models forCoordination
24 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Partial global planning
Partial global planning (PGP): exchange information to reachcommon conclusions about problem-solving process
Partial ⇒ individual agents don't generate plan for entireproblem
Global ⇒ agents use information obtained from others toachieve non-local view of problem
Three iterated stages:1 Agents deliberate locally and generate short-term plans for
goal achievement2 They exchange information to determine where plans and
goals interact3 Agents alter local plans to better coordinate their activities
Meta-level structure guides the coordination process,dictates information exchange activities
25 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Partial global planning
Central data structure is the partial global plan containing:
Objective: larger goal of the system
Activity maps: describe what agents are doing and theresults of these activities
Solution construction graph: describes how agents shouldinteract and exchange information to achieve larger goal
26 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
PGP extension
The Partial Global Planning framework was extended to theGeneralized PGP (GPGP), which introduces �ve techniquesfor coordinating activities, i.e. strategies for:
updating non-local viewpoints (share all/no/some inform.)
communicating results
handling simple (action) redundancy
handling hard (�negative�) coordination relationships(mainly by means of rescheduling)
handling soft (�positive�) coordination relationships(rescheduling whenever possible, but not �mission critical�)
27 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Joint intentions
Discussed intentions in practical (single-agent) reasoning:
Intentions also provide stability and predictabilitynecessary for social interaction⇒ intentions also signi�cant for coordination, especiallyteamwork
Helps to distinguish between non-cooperative and
cooperative coordinated activity
Basic question:In which way are individual intentions di�erent from (and whatrole do they play in) collective intentions?
Remember Cohen and Levesque's theory of intentions?They extended it to teamwork situations, introducing a notionof �responsibility�
28 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Joint intentions: example
Imagine: We try to lift a stone together, and I discover itwon't work⇒ individually rational behavior: drop the stone!However, this is not really cooperative (we should at leastinform other)Two important notions:
1 commitments (pledges or promises to underpin anintention)
2 conventions (mechanisms for monitoring commitment,mechanics of adopting/abandoning commitments)
Agents commit themselves to actions or states of a�airsCommitments are persistent, i.e. they are not droppedunless special circumstances ariseConventions de�ne these circumstances, e.g., motivationfor goal is no longer present, or it is / can never beachieved
29 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Teamwork-based model of CDPS
Teamwork-based model of CDPS (Wooldridge & Jennings,1994, 1999): a practical model of how CDPS can operate usinga teamwork approach
Stage 1: Recognition of a goal that can be achievedthrough cooperation (e.g. an agent can't do it (e�ciently)on his own)Stage 2: Team formation, i.e. assistance solicitation
if successful, this results in nominal commitment tocollective actiondeliberation phase, ends in agreement on ends (not onmeans)rationality plays a role in deciding whether to form a group
Stage 3: Plan formation involves joint means-endsreasoning, e.g. through negotiation or argumentation
Stage 4: Team action plan of joint action is executed byagents based on mutual convention
30 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Mutual modeling
Basic idea: Putting oneself in the shoes of the other.
Involves modeling others' beliefs, desires, and intentions
. . . and coordinating own actions depending on resultingpredictions
Explicit communication unnecessary
MACE one of the �rst systems (mid 1980s) to useacquaintance models for this purpose
Acquaintance knowledge involves information about othersName unique to every agentClass of group the agent belongs toRoles played by an agent in a classSkills as the capabilities of the modeled agentGoals that the modeled agent wants to achievePlans as the agent's means to achieve an end
Agent explicitly models itself!31 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Norms and social laws
Norms are established patterns of expected behavior, sociallaws often add some authority to that (can be enforced or not)⇒ Balance autonomy with goals of entire society
Features of norms and social laws in terms of conventions forMAS:
conventions make decision making easier for agents
conventions can be designed o�ine (simpler) or emerge
from within the system (more �exible)
Hard to predict at design time which norm will be optimal
Also hard to derive global conventions from agents' localpoint of view
32 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Emergent social norms and laws I
Example: the t-shirt game
agents wear red or blue t-shirt (initially at random), goal is foreveryone to wear the same color
agents are randomly paired in each round of the game, get tosee other's t-shirt color, and then may decide to switch color
Problem: agent must decide which convention to adoptalthough no global information is availablePossible update functions (=decision rules based on history):
Simple majority: agent chooses color observed most often
Simple majority with agent types: agents con�de in certain otheragents and exchange memory with them to inform their decision
Simple majority with communication on success: agents willcommunicate (successful part of) memory if success rate exceeds athreshold
Highest cumulative reward: uses strategy that has had the highestcumulative reward so far
33 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Emergent social norms and laws II
Properties of update functions:
All update functions converged to some convention
Measure: time taken to converge
Memory restarts were investigated to model �new ideas�
But also stability important (we don't want society tochange conventions all the time)
Basic result: for highest cumulative reward rule, for any0 ≤ ε ≤ 1 agents will reach agreement within n roundswith probability 1− εAlso, once reached, the convention will be stable
Convention is e�cient, i.e. it guarantees payo� no worsethan that obtainable from sticking to initial choice
Note that change of norm may be expensive in practice!
34 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
O�ine design
O�ine design is closely related to mechanism designFormally, remembering our agent model Ag : RE → Ac we cande�ne constraints 〈E′, α〉 where:
E′ ⊆ E, a constraint on the environment
α ∈ Ac, actions α as before
A social law is a set of such constraints ⇒ agents/plans arelegal if they never attempt to perform forbidden actions
useful social law problem
Given a set F ⊆ E of focal states (states that should alwaysbe allowed), a �useful social law problem� is to �nd a social lawthat allows agents to legally visit any state in F .
General problem NP-complete, tractable cases �do not appearto correspond to useful real-world cases.�
35 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Summary
Cooperative Distributed Problem Solving (CDPS)
Coherence & Coordination
Task sharing: Contract Net & Blackboard
Coordination
Partial global planning: PGP & GPGP
Joint intention: commitments & conventions
Mutual modeling: Norms, social laws, & conventions
⇒ next time: Multiagent Interactions
36 / 37
MultiagentSystems
B. Nebel,C. Becker-Asano,S. Wöl�
Motivation
Task sharing
Coordination
Models forCoordination
Partial globalplanning
Jointintentions
Mutualmodeling
Thanks
Acknowledgments
These lecture slides are based on the following resources:
Dr. Michael Rovatsos, The University of Edinburghhttp://www.inf.ed.ac.uk/teaching/courses/abs/
abs-timetable.html
Michael Wooldridge: An Introduction to MultiAgent
Systems, John Wiley & Sons, 2nd edition 2009.
Corkill, Daniel D. �Blackboard systems.� AI expert 6.9(1991): 40-47.
37 / 37